Task-Oriented Semantic Communication for AUV Collaborative Sensing with Confidence-Prioritized Triplet Compression

Autonomous Underwater Vehicles (AUVs) play an increasingly important role in marine exploration, infrastructure inspection, and environmental monitoring. In multi-AUV collaborative sensing, however, underwater communication is often constrained by limited acoustic bandwidth, redundant cross-view observations, and the lack of task-aware semantic prioritization, which together reduce communication efficiency and relevance. To address these challenges, we propose a personalized task-oriented semantic communication framework for multi-AUV systems. The framework first employs a CLIP-enhanced transformer-based scene graph generator to extract semantic triplets from underwater images. It then introduces a Personalized Semantic-Aware Ranking based on Formal Concept Analysis (PSAR-FCA) module to model user intent and prioritize task-relevant semantics. Finally, a Variational Distributed Deterministic Information Bottleneck with Confidence-Priority (VDDIB-CP) module is developed to adaptively compress and transmit high-utility semantic content under stringent bitrate constraints. Extensive experiments on two underwater datasets demonstrate the effectiveness of the proposed framework. PSAR-FCA outperforms state-of-the-art semantic summarization methods by up to 27.5% in NDCG@5, while VDDIB-CP improves overall utility by 10% and reduces transmission delay by more than 60% under constrained bandwidth. These results verify the potential of the proposed framework for efficient and task-aware semantic communication in multi-AUV underwater sensing.